mat2.h 1012 B

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748
  1. /*
  2. * Copyright (c), Recep Aslantas.
  3. *
  4. * MIT License (MIT), http://opensource.org/licenses/MIT
  5. * Full license can be found in the LICENSE file
  6. */
  7. #ifndef cglm_mat2_sse_h
  8. #define cglm_mat2_sse_h
  9. #if defined( __SSE__ ) || defined( __SSE2__ )
  10. #include "../../common.h"
  11. #include "../intrin.h"
  12. CGLM_INLINE
  13. void
  14. glm_mat2_mul_sse2(mat2 m1, mat2 m2, mat2 dest) {
  15. __m128 x0, x1, x2, x3, x4;
  16. x1 = glmm_load(m1[0]); /* d c b a */
  17. x2 = glmm_load(m2[0]); /* h g f e */
  18. x3 = glmm_shuff1(x2, 2, 2, 0, 0);
  19. x4 = glmm_shuff1(x2, 3, 3, 1, 1);
  20. x0 = _mm_movelh_ps(x1, x1);
  21. x2 = _mm_movehl_ps(x1, x1);
  22. /*
  23. dest[0][0] = a * e + c * f;
  24. dest[0][1] = b * e + d * f;
  25. dest[1][0] = a * g + c * h;
  26. dest[1][1] = b * g + d * h;
  27. */
  28. x0 = glmm_fmadd(x0, x3, _mm_mul_ps(x2, x4));
  29. glmm_store(dest[0], x0);
  30. }
  31. CGLM_INLINE
  32. void
  33. glm_mat2_transp_sse2(mat2 m, mat2 dest) {
  34. /* d c b a */
  35. /* d b c a */
  36. glmm_store(dest[0], glmm_shuff1(glmm_load(m[0]), 3, 1, 2, 0));
  37. }
  38. #endif
  39. #endif /* cglm_mat2_sse_h */